Generally, there are many methods to categorize unknown data in statistics. In many of these methods, we need sample data to determine a border of groups to which these data belong. Moreover, neural networks are also applicable to classify unknown data. Also in leaning process of neural networks, we have to prepare so-called teaching signals, i.e. sample data.
In this paper, we propose an empirical scheme to organize neural networks for clustering unknown data which belong to certain two groups. In our scheme, a neural network that satisfies an evaluation function without teaching signals are organized. This evaluation function are determined by a histogram of outputs of the neural network. Generally, neural networks map input data distribution to output one. Maximizing the evaluation function means to separate these two output distributions each other. As an organizing mechanism, genetic algorithm is used because of its convergence ability to global maximum. Some numerical results are presented to confirm a feasibility of the scheme.